[XPU] [CI] Add comprehensive tests for refactored get_padding_offset kernel

[devin-ai-integration]

Motivation

PR PaddlePaddle#7029 refactored get_padding_offset from two separate XPU kernels into a single kernel. The existing test only covers one basic scenario (seq_lens=[4,3,6], max_len=10). This PR adds comprehensive test coverage to verify correctness across edge cases and confirm behavioral equivalence with the old implementation.

Modifications

Added tests/operators/test_get_padding_offset_refactor.py with 23 test cases:

Core algorithm tests:

• Basic case matching the commit's test data (seq_lens=[4,3,6])
• Single batch (bs=1)
• Single token total (bs=1, seq_len=1)
• All sequences at full length (no padding)
• All sequences at length 1 (maximum padding)
• Mixed/identical/increasing/decreasing sequence lengths
• Large batch size (bs=64)
• Large int64 token values (overflow check)
• max_len=1 edge case
• Zero token_num edge case (all seq_lens=0)

Equivalence verification:

• Confirms x_remove_padding, cum_offsets_out, cu_seqlens_q, cu_seqlens_k match the old two-kernel implementation
• 50 randomized configurations tested for equivalence

Property-based invariant tests:

• Output array lengths match expected dimensions (100 random configs)
• cu_seqlens is monotonically non-decreasing
• batch_id_per_token values are in [0, bs-1]
• cum_offsets_out values are non-negative

Usage or Command

python -m pytest tests/operators/test_get_padding_offset_refactor.py -v --noconftest

Accuracy Tests

All 23 tests pass locally. The tests validate that:

1. The refactored single-kernel algorithm produces identical x_remove_padding, cum_offsets_out, cu_seqlens_q, cu_seqlens_k as the old two-kernel implementation
2. The new batch_id_per_token output correctly stores batch indices (replacing the old padding_offset which stored cumulative offsets)

Checklist

• Add at least a tag in the PR title.
• Format your code, run pre-commit before commit.
• Add unit tests. This PR is specifically adding unit tests.
• Provide accuracy results. All 23 tests pass.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch. (N/A - targeting develop)

Link to Devin session: https://app.devin.ai/sessions/0a98367efe304de197b296c0940bb3d4
Requested by: @ShaneGZhu

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring